Q. Why is Baidu sharing its secret self-driving sauce? A. To help China corner the market Chinese tech powerhouse Baidu plans to "open up" its self-driving car designs to help automakers in China overtake Western rivals. Big corps around the world are racing to the finish line of building and selling the first proper autonomous rides. Competition is fierce – in California, 30 organizations including Tesla, Waymo, …
Well that's not a surprise these days, everyone is doing that and customer convenience comes second in the list of reasons why.
On the other hand, in this particular arena it is rather justified. It makes sense to have a central server be aware of where everyone is and what direction they are going - makes for better traffic flow regulation if nothing else. We'll just have to find out the hard way what other data they are skimming at the same time.
Cloud: Central point of failure
It has its uses in development as data from test vehicles can be sent to the cloud for analysis, but the idea of real time cloud use to know location / speed of other vehicles dependent on cloud connectivity is not conducive to safety.
A good self driving car needs to be autonomous, not cloud dependent.
Ignoring network failures, anything needing internet will be totally screwed in huge swathes of Scotland as an example as zilch network coverage and satellite is slow data rate and v. expensive.
"Cloud: Central point of failure"
Why do you think so? This isn't an old client-server environment where the server fails and everything goes down. "Cloud" done properly means data and processing are both shared across multiple data centres, and replicated on 'disaster recovery' data centres. So it would continue to work if a whole data center goes down never mind just a single server.
The internet itself is designed to work around failures using multiple potential paths from data to flow, so network isn't a single point of failure (granted, if you need guaranteed low latency, that would be affected by network problems). And with regard to the endpoint connectivity, can be done by mobile data, initially in highly populated areas where 4G (eventually 5G?) coverage is dense, but once the technology got going that mobile internet infrastructure would quickly be deployed along all major roads. Again, if you lose one (or a few) masts/antennas, there will still be others that can pick up the slack.
What's more important is that the cars have a high enough degree of autonomy that they can anyway operate safely without needing to be cloud-connected, and the cloud connection just helps improve performance and optimise traffic.
It goes without saying, of course, that the security has got to be LOTS better than that on current cars
@jmch
"Cloud" done properly means data and processing are both shared across multiple data centres, and replicated on 'disaster recovery' data centres.
and just who is interested in doing cloud properly?
people (especially management types & aspirational techies) who hear cloud expect their single instance running in a single region to be instantly immune from any outage whether that be network outage or service failure without understanding it takes a lot more than just hosting in app in a cloud DC for the app to be resiliently available.
"This isn't an old client-server environment where the server fails and everything goes down"
client-server is easy to make highly available and resilient, make it session less, put some load balancers in, get creative with DNS. have the same solution running in multiple geographically diverse DC's, replicating the data and then give methodology a catchy title, so those above can sell the spend, like "cloud".
I'm sure Baidu know what they are talking about, not sure others do.
@jmch. Don't live in rural OZ do you ? travel 40 km from a large town and its maybe level connectivity from one dominant telco, probably not from the other excuses for comms. Granted it is a small market, but devices that demand connectivity wont work in much of the world. Biggest issue for me is do you want a vehicle that can be remotely managed ? If it can, the Bad Guys will use that control backdoor.
It doesn't have to be one or the other i.e. all local or all cloud processing, so you're both right.
The car obviously needs to be able to perform all navigation, collision avoidance, etc. autonomously, but when available it makes sense to share it's information with the cloud, so for e.g. it could receive information about conditions such as congestion on it's intended route, and perhaps send local driving conditions in return.
Baidu has a pretty decent track record as Good Guys in open source software (inasmuch as one can generalise about any company that big). The language barrier reduces their visibility to Reg readers, but I've found them extremely helpful in reaching out across it when I had occasion to work with a project where they were lead developers. This looks like another application of that principle of openness.
Are these self-driving cars going to be uploading any part of the video feed with a Windows-style "telemetry" excuse? Would such data be able to be aggregated across many self-driving cars to identify and track other cars on the road, either through their license plates or tagging and tracking other vehicles while they're in your field of view, and using the cloud part to fill in for discontinuities ("white van was tracked until point A, then lost; another sighting at point B is consistent with being the same white van, ...").
An autonomous driving system obviously needs to be able to be aware of other vehicles and make sure that it doesn't forget that they could still be around (temporarily in a blind spot or obstructed from view) before making manoeuvring decisions, but surely there are issues if these data are being aggregated in near-real time in a cloud somewhere.
Also, why stop at tracking vehicles? Surely it would need to have awareness of pedestrians, too. Maybe the current crop of cars are more suited to highway driving or driving somewhere like the US, where vehicles have right of way, so ignoring pedestrians who might suddenly walk out in front of you might make sense (until you have to go into collision-avoidance mode). For city driving, though, and in countries where jaywalking isn't a crime, surely they'll have to follow the same rules as for human drivers. Part of that is being able to figure out where pedestrians are or might be, and reading their intent, at least to some degree. Obviously, simple things like seeing that they're very close to the kerb or partly on the road is a good sign that they're looking for an opportunity to cross(*), and that can be handled by simple physical rules based on distances/location. However, reading intent is often much more complicated. If you want computers to be as good as humans, you're going to have to include things like how they act (do they turn around to look at the road as they approach a crossing, or look up at a traffic light), and figuring out where their attention is directed.
All this analysis of pedestrians (and tracking them, obviously) probably won't make it into first-generation cars, so in the initial (training) stages at least, manufacturers are going to be slurping a lot of so-called telemetry data. You can't say that blurring faces or whatever is a solution because they will need facial features to do things like gaze tracking or to judge how aware the person might be of traffic (or, eg, they're talking on a mobile phone or texting rather than paying attention to other things). The easiest thing is just to slurp everything they can, but if real-time tracking is the norm from the outset, it's hard to see how spy agencies or whatever (or even just traffic police) wouldn't want to tap into that and make sure that they continue to be able to use the system even after the AI part has been trained and downloaded as a set of real-time rules that can run on the car.
I'm sure that these sorts of concerns would definitely be looked at in Europe or the US, but in China? Somehow, I don't think so.
* another, unrelated scenario with self-drive cars strikes me. If you're coming to a crossing and you see someone that you know (or think you know) and make eye contact and give them a nod or something, how are they going to interpret that? Maybe they don't know that you're not the driver. If you were the driver, they could take your gesture or general demeanour as giving them the OK to cross the road in front of you. The car's not going to understand that...
If I generate an image of a traffic light gone red in front of an autonomous vehicle, will it stop?
When all cars are autonomous, will the insurance companies resort to hacking vehicles to cause accidents in order to drum up business?
When will we have a CPU that I can't break into by being on the same system where it runs?
Why does pizza taste so good?
What are male nipples good for?
Why can't a blockchain Quantum AI Pubic Watson Cloud solve this problem?
Please make these systems so sensitive that I can look at a car funny and it will run off the road or stop in the middle of traffic! Good times!
"Officer, those kids waved their arms in the air at my car and it stopped in traffic and caused this pile-up of self-driving crap-machines full of self-hating over-compensating ex-Uber drivers, and I spilled my latte on my Dockers."
I can't help thinking that if this was a Western company, one would not spend so much ink pointing out how there must be ulterior motives for their openness (never mind that they did say it clearly: we want to help Chinese manufacturers compete with the bigger-budgeted shops). Take a look at the coverage of Tesla's patent sharing announcement for example.
This post has been deleted by its author
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
